ocr

OCR UV Scripts Part of uv-scripts — self-contained UV scripts you run on Hugging Face Jobs in one command. A model zoo of OCR scripts — one per model — that add a `markdown` column to an image dataset. Pick a model from the table below, point it at your dataset, and run it on a GPU with one command. A few recipes do **structured extraction** instead — image *or* text → JSON given a schema (see Structured extraction below). Two more companions sit alongside: `pp-doclayout.py` detects layout regions (bboxes for text/title/table/figure/…) instead of text, and `ocr-vllm-judge.py` compares model outputs head-to-head. Quick Start Run OCR on any dataset without needing your own GPU: This will: Process the first 10 images from your dataset Add OCR results as a new `markdown` column Push the results to a new dataset View results at: `https://huggingface.co/datasets/[your-output-dataset]` Serve a model as a live endpoint The recipes here run as batch jobs. Some models also have a **`-server.py` sibling recipe** that runs the same dataset→dataset batch job through an in-job `vllm serve` + concurrent driver — measurably faster (continuous batching stays fed) and more robust (one bad image fails one request, not a whole batch); see SERVING.md for the architecture, A/B numbers, and which models officially document server mode. To call a model interactively, from an agent, or with concurrent ad-hoc requests, you can instead run it as a temporary endpoint: HF Jobs serving exposes a port on a GPU Job, giving an OpenAI-compatible endpoint that runs until the job is cancelled or its `--timeout` is reached. See serving-unlimited-ocr.md for a worked example serving Baidu's Unlimited-OCR — with vLLM (official image) or SGLang. To OCR a whole corpus of single-page images instead, the batch recipe `unlimited-ocr-vllm.py` is the better fit (it's single-image only). **Multi-page** documents need a server: both vLLM and SGLang read clean multi-page docs, but **SGLang is the more robust** — on